Posted this on LinkedIn first: a response to the unveiling of GPT-5.

 

'Reading the abstract (Chain of Thought reasoning is “a brittle mirage that vanishes when it is pushed beyond training distributions”) practically gave me deja vu. In 1998 I wrote that “universals are pervasive in language and reasoning” but showed experimentally that neural networks of that era could not reliably “extend universals outside [a] training space of examples”.

The ASU team showed that exactly the same thing was true even in the latest, greatest models. Throw in every gadget invented since 1998, and the Achilles’ Heel I identified then still remains. That’s startling. Even I didn’t expect that.

And, crucially, the failure to generalize adequately outside distribution tells us why all the dozens of shots on goal at building “GPT-5 level models” keep missing their target. It’s not an accident. That failing is principled.'

And the principle is far older than LLMs: it goes back to the AI wars of the 60s and 70s. ML-based AI was a mystery religion that produced miracles that could not be explained. The miracles were flashy enough to get the plodding tortoises of symbolic logic and linguistics out of Big AI (universities and tech bro startups) and banish them to the margins. Gary Marcus, who wrote the critique below, was one of the survivors.

"In his first book, The Algebraic Mind (2001), Marcus challenged the idea that the mind might consist of largely undifferentiated neural networks. He argued that understanding the mind would require integrating connectionism with classical ideas about symbol-manipulation."

Gary Marcus Wikipedia entry

 

GPT-5: Overdue, overhyped and underwhelming. And that’s not the worst of it.